@nicconley: This model was trained on 170 years of human video and zero hours of robot data Meet Dyna 2 built by Dyna Robotics in Redwood City It was trained on over 1000000 hours of video shot from peoples own point of view which works out to about 170 years of continuous waking experience None of it came from robots and thats where the bottleneck disappears Robots normally learn from teleoperation data where a human manually drives the machine through a task Its slow expensive and there isnt much of it Video is everywhere Dyna 2 generates the future video frames and the motor commands together and Dyna argues thats what gives it a sense of how objects behave before it moves Heres where it gets insane Nvidia ran this experiment in April on 20000 hours and called it the first scaling law for robot dexterity Dyna says it ran the same thing at 50 times that scale Its performance kept climbing the whole way with no plateau But the craziest part 13 minutes of data was enough to teach a pair of five fingered hands to twist open a bottle cap If you want to stay updated with the latest tech news follow for more Hashtags: #Dyna2 #DynaRobotics #Dyna2Model #DynaRoboticsRedwoodCity #Dyna2Trained
